- Building
pileGo! is for people selling their own secondhand clothes. You snap the pile once, and pileGo lists each piece on Depop, Poshmark, Mercari, Vinted and eBay. When something sells on one site, it comes down from all the others, so you never sell the same hoodie to two people.

The posting runs on HarnessRouter. Every seller gets their own hosted browser, and our pileGo listing publisher Harness works in it: it posts listings, takes them down, changes prices and checks for sales. You sign in to each marketplace once, inside pileGo. After that there's nothing to install, it works from your phone, and it keeps going after you close your laptop.
What we shipped, Nov 15 to 18
- One agent for four marketplaces. Depop, Poshmark, Mercari and Vinted all run through the Harness. eBay uses its official API.
- A live view of every listing going up. There's more on this under Loved.
- Batch scanning. Shoot a whole pile, tap Finish, and every listing gets written and posted together.
- Demo accounts. They make one real Vinted post and simulate the other three sites, so we can show pileGo off without filling real marketplaces with test listings.
- 10 bugs found and fixed. Each one is written up with steps to reproduce in our build log.
- Blocked
1. Running out of credit looked like a normal failure
What happened: when our model key was refused or our credit ran out, every job failed, and our worker retried each one three times. All three attempts were doomed, and the seller saw their listing fail again and again.
To reproduce
- Start a background task on a Harness while the model key is refused or the balance is empty.
- Read the finished task. It looks like any other failed task. The only sign of the real reason is the wording of the error message.
What we did: we now search the error text for phrases like "TokenRouter key was refused" and "no_credit". When we find one, we pause the job for 15 minutes instead of retrying. That works, but it breaks the moment the wording changes. (See Want #1.)
2. Our browser host's free minutes ran out mid-hackathon
What happened: a seller's sign-in window closed halfway through signing in, and after that every new browser was refused. We had to add a second browser host (Steel) in the middle of the hackathon to keep going.
To reproduce: use up a free plan's browser minutes while a live sign-in window is open.
3. "Continue with Google" looked broken
What happened: Google and Apple sign-in open in a popup tab, and the live view showed only the first tab. Sellers pressed the button and nothing seemed to happen. We now switch the view to whichever tab opened last.
4. Vinted refused every sign-in from a cloud browser
What happened: even through home internet connections, Vinted answered every cloud sign-in with "Your session has been blocked". Our workaround lets sellers bring the sign-in over from the browser they already use on Vinted.
- Loved
- Answers we can trust. The guide pointed us at base
dshonclaude-sonnet-5.5for jobs where the answer has to be one strict JSON object, and it was right. Every job ends with a clear result, such as listed, needs_sign_in or needs_seller, and the app acts on it directly. - Each agent can only reach one seller. HarnessRouter passes our per-job token through to our tools. That ties each task to one seller's browser and one item, so an agent can never touch anyone else's account. This is what made us comfortable letting an agent work in real people's accounts.
- No double posts. We start every task with an idempotency key, so a retried start can never post the same item twice.
- Built-in guardrails. A cap of 120 steps, a 25-minute timeout and a $2 cost limit per job mean a confused agent can't run up a bill.
- Memory we could switch off. With it off, one seller's items never carry over into another seller's job.
- Progress updates the seller can see. The agent reports each step as it goes, and we turned those reports into one pixel-art lane per marketplace. The lanes only move when the agent really finishes a step.

- Answers we can trust. The guide pointed us at base
- Want
1. A clear error code for "out of credit", marked as "don't retry"
Problem: a refused key or an empty balance looks like any other failure, so we had to match on the English wording of the error message (Blocked #1).
Who it affects: any team that runs background tasks with automatic retries.
How it would help: developers stop paying for retries that can't succeed, and their code doesn't break when the wording changes. End users get an honest "back in 15 minutes" instead of a pile of failed attempts.
2. A webhook when a background task finishes
Problem: we check every running task once a minute to see if it's done. A listing can sit finished for up to a minute before the seller sees it go live, and most of those checks find nothing new. (If this already exists, we missed it. Please point us to it!)
Who it affects: any app that shows people live agent progress.
How it would help: results show up the moment they happen, and developers can remove a whole polling loop.
3. A built-in "dry run" for tasks
Problem: we test against real marketplace accounts. To be sure a test never posts anything, we had to write dry-run rules into our instructions and also block it in our own tools.
Who it affects: any team whose agent acts on real accounts, such as posting, paying or deleting.
How it would help: a task-level flag that blocks any tool marked as a write would make testing safe by default, with no hand-rolled safety rules. Teams could test against real accounts without risking a live post.
Loading